Commodity Image Retrieval Method Based on CNN and Score Fusion

Abstract

Content-based image retrieval (CBIR) is an important research direction of computer vision and has been well studied for a long period. The important aim in CBIR is to reduce the semantic gap issue that improves the performance of image retrieval. It plays an important role in the field of e-commerce. This paper proposes a CBIR method based on ResNet and EfficientNet, and conducts experiments on the dataset of eProduct Visual Search Challenge 2021. First, because the dataset lacks direct query-index tags and the categories are not balanced, transfer learning is used to fine-tune the model with a balanced sub-dataset constructed by data augmentation. Second, the semantic features are extracted by the ResNet and EfficientNet and similarity comparison is calculated between the query and index image. Finally, the similarity result is decided through fusing the scores obtained by the two CNNs through a strategy. Experiments on the eProduct dataset demonstrate our algorithm can achieve a good performance.